Perceptually-based objective measure for non-intrusive speech quality assessment

This paper proposes a new perceptuallybased method for assessing speech quality and evaluates its performance. The method is based on comparing the received speech to an appropriate reference representing the closest match from a preformulated codebook. The codebook holds a number of optimally clustered speech parameter vectors extracted from a large number of various undistorted clean speech records. The objective auditory distances between vectors of the distorted speech signal and their corresponding matching references are then measured and appropriately converted into an equivalent subjective score. The optimal clustering of the reference codebook is achieved by using a dynamic k-means method. Efficient data mining technique known as Self-Organising Map is used to match the distorted speech vectors to the references. Speech parameters derived from Bark spectrum analysis, and Mel-Frequency Cepstral coefficients (MFCC) are used to provide speaker independent parametric representation of the speech signals as required by an output-based quality measure.